Piano Concerto Source Separation Based on Speech Separation and Channel Attention

The Mossformer model excels in speech separation but has not been effectively applied to music source separation. Music sources have complex characteristics and higher sampling rates, making separation tasks more challenging. We addressed a rarely explored task of separating piano concerto recordings into individual piano and orchestral tracks. This process involves intricate coordination between the piano and orchestra, creating highly complex audio signals in both time and frequency domains. Our main contributions include: (1) adapting the speech separation model for the novel task of piano concerto source separation, constructing and processing a specialized dataset.(2) introducing channel attention in the separation module to dynamically adjust feature focus based on instrument characteristics, enhancing key features. Experiments on the Piano Concerto Dataset (PCD) showed improved separation performance, with a 0.22dB average Signal-to-Distortion Ratio (SDR) increase over the baseline model.